THE PREDICTION OF SALE TIME SERIES BY ARTIFICIAL NE URAL [611445]

THE PREDICTION OF SALE TIME SERIES BY ARTIFICIAL NE URAL
NETWORK

Abstract: The article presents one of the methods of predicti on of time series with the support of
MATLAB program when Neural Network Toolbox is used. There are mentioned the steps of
calculation. The setup of parameters plays a key ro le for correct prediction. The method of
calculation can be used for any time series.

Keywords: Prediction, time series, calculation, MATLAB, Neura l Network Toolbox

1. Introduction
The article presents the steps of calculation of pr ediction of sell time series. The Neural Network To olbox of
MATLAB environment is used. The correct prediction is very complicated task because the components of time
series in economy and financial are quite often sto chastic and random. Therefore the prediction is ver y difficult
and sometimes impossible, when the behaviour of tim e series is random. This is quite often case of tim e series
on the stock market. A little better situation it i s with sale time series. The sale time series inclu de also
deterministic components such as periodicity and/or tendency that are easy recognizable. It could be m entioned
consumption or sell of some products, for example w eek periodicity is caused by behaviour of customers during
the week, year tendency is caused by physical pheno mena, Gompertz curve is caused by saturation of mar ket,
etc. These periodicity and tendency enable reasonab le prediction. The case study is focussed on consum ption of
energy.

2. Used Methods
There are hundreds of methods for prediction of tim e series. The competitions are held where the best prediction is
evaluated. The program MATLAB with Neural Network T oolbox is a strong tool for prediction. The steps o f
predictions are as follows: download of input and o utput time series, choice of the type of neural net work, set up the
number of layers, the number of neurons, transfer f unctions and some other parameters necessary for th e calculation.
When the neural network is build up, the process of learning and testing is necessary to run. After th e calculation it is
necessary to evaluate the results of prediction. Th e parameters of neural network are possible to chan ge in case of
bad prediction. The process of learning and testing is possible repeat several times and with changed parameters.
When the prediction is correct, it could be used as a support for decision making for the setup of pro duction plan. The
results could be exported for graphical visualisati on.

3. Case study
This case study presents the prediction of consumpt ion of energy of a town, where the day and week per iodicity
is recognisable together with season tendency influ ence. The behaviour of people and weather could be
evaluated as a chaotic. The input data are availabl e at input.csv and output data for learning at target.csv file.
The data are downloaded by orders

p=(load('input.csv'))';
t=(load('target.csv'))';
nntool

and the order nntool for opening neural network is called. The import o f input time series is done by means of
menu Import p Input data Import OK and output data by means of menu Import t Target data Import OK . See
Fig.1.

Fig. 1 Import of input and output time series

The order New Create displays the screen that enables to choose the typ e of neural network, the number of layers,
the number of neurons, transfer functions and some other parameters necessary for the calculation. The process of
build-up of neural network is done by order Create . See Fig.2.

Fig. 2 The build-up of neural network

The neural network named network1 must be trained. The procedure starts by order Train-Train
Network . The process of training is presented in the windo w. After the process of training it is possible to use
options for display graphs to evaluate the performa nce. See Fig. 3.

Fig. 3 Displaying the process of training of neural netwo rk

When the neural network is trained and successfully validated, the data could be exported by the order Export –
network1 – Export OK – Close for drawing of real and predicted values of time s eries. The graph is plotted by
orders

sim(network1,p);
plot(t)
hold on
plot(ans,'g').

The graph is presented at Fig.4. The graph presents a history of 350 samples and the next 50 samples o f
prediction. The fig.5 presents zoomed graph of pred iction.

Fig. 4 Historical and predicted data of time series
Prediction

Fig. 5 Zoomed graph of time series

5. Conclusion
The sale time series usually contains recognizable tendency and periodicity trends, therefore the pred ictions
gives quite often useful information. The predictio n supports right decision when production plan is p repared.